OctoInform Beta

OpenAI preprint, September 2026 · §5. Measurable labels and finite blocks · §5

Proposition 5.2 [Measurable local labels]: For |I|≥3|I| \ge 3, one measurable set of canonical codes over the parameter space gives, almost surely, the exact lines at every slot

Exact lines of a saturated array object carry measurable labels.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

Consider stationary random arrays of states in which every state carries a parameter. Given the parameters of a tuple of states, one state for each slot, which circle-valued functions of the state at one slot are matched exactly by functions of the states at the other slots, and can these exact lines be labelled measurably by the parameter of their own slot, compatibly with products and with time?

A standard Borel space is a nonempty set with a σ\sigma-algebra under which it is isomorphic to a Polish space with its Borel sets.✦ Fix one, XX, with a measurable automorphism TT, and a finite index set II, whose elements are called slots. An array of order nn over a space YY is a function from the index tuples v∈Inv \in I^n to YY, a point of YInY^{I^n}✦, and a map f:Y→Zf\colon Y \to Z acts on arrays entrywise, as fn:YIn→ZInf_n\colon Y^{I^n} \to Z^{I^n}.✦ Arrays of order zero are points of YY✦, arrays of order one are tuples (yi)i∈I(y_i)_{i \in I}, and arrays of order two are squares (yas)a,s∈I(y_{as})_{a, s \in I}. The plane of an array of order n+1n+1 that fixes axis aa at slot ii is the array of order nn whose entry at vv is the entry at the tuple obtained by inserting ii into vv at position aa.✦

Initial array laws of TT are probability measures mnXm^X_n on XInX^{I^n}, one for each order nn, each invariant under TT acting entrywise.✦ Initial array laws mXm^X of TT are consistent if every plane of an array drawn from mn+1Xm^X_{n+1} has law mnXm^X_n, and proper if, for every subset JJ of II other than II, the planes that fix the last axis at the slots of JJ are independent under mn+1Xm^X_{n+1}, each with law mnXm^X_n.

An array object MM over initial array laws mXm^X of TT consists of a standard Borel space YY of states with an automorphism TYT_Y, a standard Borel space UU of parameters with an automorphism TUT_U, measurable maps x:Y→Xx\colon Y \to X and par:Y→U\mathrm{par}\colon Y \to U with x∘TY=T∘xx \circ T_Y = T \circ x and par∘TY=TU∘par\mathrm{par} \circ T_Y = T_U \circ \mathrm{par}, and probability measures mnm_n on YInY^{I^n} that are invariant under TYT_Y, consistent and proper in the same sense, and carried to mnXm^X_n by xx acting entrywise.✦ Write κw(n)\kappa^{(n)}_w for the conditional law under mnm_n of an array given that its parameter array, par\mathrm{par} applied entrywise, is ww; it is a Markov kernel from UInU^{I^n} to YInY^{I^n}.✦ Moreover, an array object MM is conditionally local: given the parameter array ww of an array drawn from mn+1m_{n+1}, the planes that fix the last axis at the slots jj of any J≠IJ \neq I are independent, the one at slot jj with law κ(n)\kappa^{(n)} at the plane of ww at slot jj. Also, an array object MM is locally independent: under m0m_0 the maps xx and par\mathrm{par} are independent.

The parameter law of order nn is the law of the parameter array under mnm_n: ν\nu on UU at order zero, π\pi on UIU^I at order one.✦ At order zero, write κu=κu(0)\kappa_u = \kappa^{(0)}_u for the law of a state with parameter uu, and Hu=L2(κu)H_u = L^2(\kappa_u), a complex Hilbert space; at order one, write p=(pi)i∈Ip = (p_i)_{i \in I} for a parameter tuple.

Let Λ\Lambda be a Markov kernel from UIU^I to YIY^I such that, for every p∈UIp \in U^I and j∈Ij \in I, the coordinate yjy_j has law κpj\kappa_{p_j} under Λp\Lambda_p and the coordinates yty_t, t≠jt \neq j, are independent under Λp\Lambda_p. A Markov kernel Λ\Lambda from UIU^I to YIY^I is a version of κ(1)\kappa^{(1)} if Λp=κp(1)\Lambda_p = \kappa^{(1)}_p for π\pi-almost every p∈UIp \in U^I.

Extensions and Saturation

An extension of MM by an array object NN over the same initial laws is a pair of measurable maps, σ:YN→YM\sigma\colon Y_N \to Y_M between the state spaces and ρ:UN→UM\rho\colon U_N \to U_M between the parameter spaces, each intertwining the automorphisms, with xM∘σ=xNx_M \circ \sigma = x_N and parM∘σ=ρ∘parN\mathrm{par}_M \circ \sigma = \rho \circ \mathrm{par}_N, such that σn\sigma_n carries mnNm^N_n to mnMm^M_n for every nn.✦ The pullback of an extension (σ,ρ)(\sigma, \rho) at order nn is the linear isometry F↦F∘σnF \mapsto F \circ \sigma_n from L2(mnM)L^2(m^M_n) to L2(mnN)L^2(m^N_n).✦

A data field of an object is named by its kind dd: the parameter field 𝒫n+1\mathcal{P}_{n+1}, or for a slot jj the outside field 𝒪n+1,j\mathcal{O}_{n+1,j} or the enlarged field 𝒢n+1,j\mathcal{G}_{n+1,j}, each a σ\sigma-algebra on the arrays YIn+1Y^{I^{n+1}} of order n+1n+1. The projection EdME^M_d is the conditional expectation onto the data field of kind dd of MM, the orthogonal projection of L2(mn+1M)L^2(m^M_{n+1}) onto the functions measurable for that field.✦ The object MM is saturated if these projections commute with every pullback: EdN(F∘σn+1)=(EdMF)∘σn+1E^N_d(F \circ \sigma_{n+1}) = (E^M_d F) \circ \sigma_{n+1}(1)✦ for every extension (σ,ρ)(\sigma, \rho) of MM by an array object NN, every order nn, every kind dd and every F∈L2(mn+1M)F \in L^2(m^M_{n+1}).

Hilbert Codes

Fix a countable family of sets generating the σ\sigma-algebra of YY, and take the algebra of sets that it generates.✦ The test functions are the finite sums ∑lcl𝟏Al\sum_l c_l \mathbf{1}_{A_l} with coefficients cl∈ℚ(i)c_l \in \mathbb{Q}(\mathrm{i}) and sets AlA_l in that algebra✦; an enumeration of them, fixed once and for all, gives every finite measure μ\mu on YY its test sequence (tn)n∈ℕ(t_n)_{n \in \mathbb{N}} in L2(μ)L^2(\mu)✦.

A Hilbert code for κ\kappa is a pair q=(u,z)q = (u, z) of a base point u∈Uu \in U and a real sequence zz that is a distance code for the test sequence (tn)(t_n) of κu\kappa_u: for all indices nn and mm, zn≥0,zn≤∥tn−tm∥+zm,∥tn−tm∥≤zn+zm,infkzk=0.z_n \ge 0, \qquad z_n \le \|t_n - t_m\| + z_m, \qquad \|t_n - t_m\| \le z_n + z_m, \qquad \inf_k z_k = 0.(2)✦ The decoded vector of qq is a vector vv of HuH_u with ∥v−tn∥=zn\|v - t_n\| = z_n for every nn✦, and the code of a vector v∈Huv \in H_u is (u,(∥v−tn∥)n)(u, (\|v - t_n\|)_n).✦ The representative of qq is the function q̃\tilde q on YY whose value at yy is the limit, as m→∞m \to \infty, of tnm(y)t_{n_m}(y), the test function read pointwise, where nmn_m is the least index with znm<2−mz_{n_m} < 2^{-m}.✦

The distance dist⁡(q,r)\operatorname{dist}(q, r) between two codes is the distance in L2(κu)L^2(\kappa_u) between their representatives, uu the base point of qq.✦ A code qq with base point uu is unimodular if |q̃|=1|\tilde q| = 1 holds κu\kappa_u-almost everywhere✦, and canonical if its decoded vector vv has a positive real inner product ⟨tn,v⟩\langle t_n, v \rangle at the first index nn for which ⟨tn,v⟩≠0\langle t_n, v \rangle \neq 0, a condition that fixes the phase of vv.✦

Exact Lines and Labels

Let p∈UIp \in U^I be a parameter tuple. A tuple (qi)i∈I(q_i)_{i \in I} of canonical unimodular codes, qiq_i with base point pip_i, is exact over pp if |∫YI∏i∈Iq̃i(yi)dΛp(y)|=1,\Bigl| \int_{Y^I} \prod_{i \in I} \tilde q_i(y_i) \, d\Lambda_p(y) \Bigr| = 1,(3)✦ and ℰ(p)\mathcal{E}(p) is the set of exact tuples over pp. The exact lines at slot ii over pp are the codes found at slot ii of an exact tuple: ℰi(p)={qi:(qj)j∈I∈ℰ(p)}.\mathcal{E}_i(p) = \{\, q_i : (q_j)_{j \in I} \in \mathcal{E}(p) \,\}.(4)✦

A parameter square e∈UI×Ie \in U^{I \times I} has rows ea⋅=(eas)s∈Ie_{a\cdot} = (e_{as})_{s \in I}, the planes fixing the first axis, and columns e⋅s=(eas)a∈Ie_{\cdot s} = (e_{as})_{a \in I}, the planes fixing the last axis. An exact square over ee is an array (qas)a,s∈I(q_{as})_{a, s \in I} of codes whose row (qas)s∈I(q_{as})_{s \in I} is exact over ea⋅e_{a\cdot} for every aa and whose column (qas)a∈I(q_{as})_{a \in I} is exact over e⋅se_{\cdot s} for every ss; 𝒬(e)\mathcal{Q}(e) is the set of exact squares over ee.✦

A label is a pair (u,q)(u, q) of a parameter u∈Uu \in U and a Hilbert code qq for κ\kappa.✦ For a set SS of labels, the fibre over uu is the set SuS_u of codes qq with (u,q)∈S(u, q) \in S. A set SS of labels is separated if it is measurable, every (u,q)∈S(u, q) \in S has qq canonical and unimodular with base point uu, and any two distinct codes q,rq, r in one fibre satisfy dist⁡(q,r)≥1\operatorname{dist}(q, r) \ge 1.

Projective Classes and Time

The group of measurable circle-valued maps on YY is the group, under pointwise multiplication, of measurable maps from YY to the unit circle with the Borel structure it inherits from ℂ\mathbb{C}.✦ For a measure μ\mu on YY, the maps equal μ\mu-almost everywhere to a constant form its subgroup of constant phases✦, and the projective circle group of μ\mu is the quotient: circle-valued maps modulo constant phases and changes on null sets.✦ The circle representative qˆ\hat q of a code qq is its representative read as a point of the circle at each yy.✦ For u∈Uu \in U, the projective class [q][q] of qq is the class of qˆ\hat q in the projective circle group of κu\kappa_u✦, and for a set SS of labels the local projective group GS(u)G_S(u) is the subgroup of that group generated by the classes [q][q], q∈Suq \in S_u.✦

For a code qq with base point uu, let f=qˆ∘TY−1f = \hat q \circ T_Y^{-1}, a vector of HTUuH_{T_U u}, and let cc be its first nonzero inner product ⟨tn,f⟩\langle t_n, f \rangle with the test sequence of κTUu\kappa_{T_U u}. The time transport θ(q)\theta(q) of qq is the code of the vector (|c|/c)f(|c| / c)\, f of HTUuH_{T_U u}.✦

The Claim

Theorem 1. Let MM be an array object over initial array laws of TT with |I|≥3|I| \ge 3, and let Λ\Lambda be a Markov kernel from UIU^I to YIY^I with the marginals κpj\kappa_{p_j} and the independence off each slot described in the setting. In (i) to (v) suppose moreover that MM is saturated and that Λ\Lambda is a version of κ(1)\kappa^{(1)}.

  1. There is a separated set SS of labels such that ℰi(p)=Spifor every i∈I, for π-almost every p∈UI.\mathcal{E}_i(p) = S_{p_i} \quad \text{for every } i \in I, \text{ for $\pi$-almost every } p \in U^I.(5)✦

  2. For each slot i∈Ii \in I there is a separated set SS of labels such that ℰi(p)=Spifor π-almost every p∈UI.\mathcal{E}_i(p) = S_{p_i} \quad \text{for $\pi$-almost every } p \in U^I.(6)✦

  3. If sets SiS^i of labels, one for each slot i∈Ii \in I, satisfy ℰi(p)=Spii\mathcal{E}_i(p) = S^i_{p_i} for π\pi-almost every pp, then for all slots aa and jj Suj⊆Suafor ν-almost every u∈U.S^j_u \subseteq S^a_u \quad \text{for $\nu$-almost every } u \in U.(7)✦

  4. For a separated set SS of labels satisfying (5) and for ν\nu-almost every u∈Uu \in U, taking projective classes is a bijection from the fibre onto the local projective group: Su→GS(u),q↦[q].S_u \longrightarrow G_S(u), \qquad q \longmapsto [q].(8)✦

  5. If a set SS of labels satisfies (5), then for ν\nu-almost every u∈Uu \in U time transport is a bijection between fibres: Su→STUu,q↦θ(q).S_u \longrightarrow S_{T_U u}, \qquad q \longmapsto \theta(q).(9)✦

  6. Let e∈UI×Ie \in U^{I \times I} and a∈Ia \in I, and suppose that every exact tuple over the row ea⋅e_{a\cdot} is row aa of some exact square over ee. Then restriction to row aa is a bijection from the exact squares over ee onto the exact tuples over ea⋅e_{a\cdot}: 𝒬(e)→ℰ(ea⋅),(qbs)b,s∈I↦(qas)s∈I.\mathcal{Q}(e) \longrightarrow \mathcal{E}(e_{a\cdot}), \qquad (q_{bs})_{b, s \in I} \longmapsto (q_{as})_{s \in I}.(10)✦

In the paper, the result reads: on a ν\nu-conull set of parameters uu there is a measurable countable family of abelian groups ℒ(u)\mathcal{L}(u) of labels of lines in HuH_u, with measurable modulus-one representatives, identity 11, orthogonal, of mean zero off the identity and with products respecting the group law, such that for π\pi-almost every pp the exact lines at slot ii are exactly ℒ(pi)\mathcal{L}(p_i) and span what the paper calls the active space, with measurable isomorphisms Φst(p)\Phi_{st}(p) for slots ss and tt, time isomorphisms τu\tau_u respecting exact matching, and, for almost every square, a label at one cell determining a unique array matched along every row and column.

The fibres SuS_u of (5) take the place of the paper’s groups ℒ(u)\mathcal{L}(u), matched with the local projective groups by (8).

What Is Checked

The claim is stated over an array object MM with at least three slots over initial array laws of TT, saturated in (i) to (v); a Markov kernel Λ\Lambda with the marginals κpj\kappa_{p_j} and independence off each slot, a version of κ(1)\kappa^{(1)} in (i) to (v); Hilbert codes for κ\kappa with their distance and the conditions of being canonical and unimodular; exact tuples, exact lines and exact squares over Λ\Lambda; the projective circle groups of the measures κu\kappa_u; and time transport of codes.

The library proves that the map σ\sigma of an extension (σ,ρ)(\sigma, \rho), acting entrywise, is measure preserving from mnNm^N_n to mnMm^M_n.✦ For circle-valued maps on YY it proves that the measurable ones, and among them those almost everywhere constant, are closed under products and inverses, that measurable ones lie in L2L^2✦, and that the circle representative of a code is measurable✦. For time transport it proves that the code of qˆ∘TY−1\hat q \circ T_Y^{-1} is unimodular✦ and that every unimodular code has a nonzero inner product with some term of the test sequence✦.

The arrays of order zero over YY form a space measurably isomorphic to YY, an array being sent to its entry at the empty tuple✦, and a measurable equivalence of YY acts entrywise as a measurable equivalence of the arrays of order nn, with inverse the entrywise action of the inverse equivalence✦.